A weather-aware traffic flow prediction method based on deep spatio-temporal disturbance modeling

Abstract

The application discloses a weather-aware traffic flow prediction method based on deep space-time disturbance modeling. The method adopts a complete architecture of "decoupling-prediction-restoration": the space-time disturbance term generated by weather on traffic flow is explicitly modeled by enhancing the disturbance network; the generated historical disturbance is subtracted from the original traffic flow to realize the decoupling of weather interference and pure traffic dynamics; the self-adaptive space-time graph convolution module is used to predict the decoupled pure traffic data; finally, the future weather disturbance is superimposed back to the prediction result to complete the restoration. The application solves the problems that the traditional model does not fully utilize weather information and the precision sharply decreases under adverse weather, and shows excellent prediction performance and strong weather robustness on the PEMS dataset, and can be widely applied to urban intelligent traffic control.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation systems (ITS) and deep learning technology, specifically to a weather-sensing traffic flow prediction method based on deep spatiotemporal disturbance modeling. Background Technology

[0002] In intelligent transportation systems, accurate short-term traffic flow prediction is crucial for traffic signal control, route planning, and congestion warning. Current mainstream methods, such as STGCN and Graph WaveNet, primarily rely on spatiotemporal graph neural networks. However, existing technologies suffer from the following shortcomings: 1. They simply stitch together weather information as auxiliary features, failing to model the nonlinear interference of weather as an "active disturbance source" on traffic flow; 2. Under severe weather conditions such as rain and snow, the lack of specialized disturbance modeling results in poor model robustness and a significant drop in prediction accuracy; 3. They rely on a fixed physical road network matrix, making it difficult to capture implicit connections between road segments; 4. They lack interpretability and cannot quantify the strength of weather-related disturbances affecting different road segments. Summary of the Invention

[0003] To address the aforementioned problems, this invention proposes a weather-aware traffic flow prediction method based on deep spatiotemporal disturbance modeling (DeepTD). The technical solution adopted by this invention is as follows: A weather-aware traffic flow prediction method based on deep spatiotemporal disturbance modeling adopts a "decoupling-prediction-reconstruction" architecture. First, the spatiotemporal disturbance of weather on traffic flow is extracted by enhancing the disturbance network; second, the disturbance term is removed from the original flow, allowing the prediction model to focus on learning pure traffic dynamics; finally, the predicted value is restored by superimposing future disturbance terms. The positive effects of this invention are: (1) Explicitly modeling weather disturbances improves the prediction accuracy of the model under severe weather conditions. (2) The robustness of the model is enhanced by the decoupling architecture, and the MAE decay rate is much lower than that of existing mainstream models. (3) The adaptive graph learning mechanism can automatically discover the implicit spatial relationships of the road network. (4) It has interpretability and can output the intensity of weather interference on each road segment, supporting refined management. Attached Figure Description

[0004] Figure 1 Traffic case analysis diagram provided for this invention (showing a comparison of traffic flow distortion on sunny days and rainy / snowy days); Figure 2 This is a schematic diagram of the overall model framework of the present invention; Figure 3 This is a comparison chart of the performance experimental results of the present invention and the comparative model; Figure 4 This is a thermal map illustrating the robustness of the present invention under severe weather conditions. Detailed Implementation

[0005] This embodiment employs the DeepTD framework and is validated on the PEMS04 and PEMS08 datasets. In Example 1, for 15-minute short-term predictions on the PEMS04 dataset, the input includes 13-dimensional weather features such as temperature, humidity, and precipitation type. The hidden dimension is set to 128, and the Chebyshev order K=3. Experimental results show a MAE of 16.52, significantly outperforming benchmark models such as STGCN. In Example 2, for 60-minute predictions on the PEMS08 dataset, this method demonstrates stronger long-term prediction stability. In Example 3, robustness is verified by selecting rainy day data; the performance degradation of this invention under strong disturbances is only 10.4%–12.4%, far superior to the 14%–20% of traditional models. This invention decouples weather noise, enabling the model to maintain high-precision flow capture capabilities even under complex weather conditions.

Claims

1. A weather-sensing traffic flow prediction method based on deep spatiotemporal perturbation modeling, characterized in that, The method adopts a "decoupling-prediction-reconstruction" architecture, with the following specific steps: Step 1: Input historical traffic flow data, weather features, and temporal features into an enhanced perturbation network to generate node-level initial perturbations; Step 2: Spatially smooth the initial perturbations using graph convolution and generate historical spatiotemporal perturbation terms through a temporal modulation mechanism; Step 3: Subtract the historical spatiotemporal perturbation terms from the original historical traffic flow data to obtain decoupled pure traffic dynamics data; Step 4: Input the pure traffic dynamics data into an adaptive spatiotemporal graph convolution module to predict the pure traffic state at future time steps; Step 5: Generate future spatiotemporal perturbation terms based on future weather forecast information and superimpose them onto the pure traffic state at the future time steps to reconstruct the final traffic flow prediction result.

2. The weather-sensing traffic flow prediction method based on deep spatiotemporal perturbation modeling according to claim 1, characterized in that: The enhanced perturbation network in step one includes an encoder, a generator + residual, and a learnable scaling factor γ·tanh-1.5, used to fit the nonlinear perturbation intensity of different road segments from multi-source weather and time features.

3. The weather-sensing traffic flow prediction method based on deep spatiotemporal perturbation modeling according to claim 1, characterized in that: The spatial smoothing process in step two uses graph convolutional networks to capture the correlation of disturbances between adjacent road segments; the temporal modulation mechanism is used to adapt to the periodic disturbance patterns of different time periods and dates.

4. The weather-sensing traffic flow prediction method based on deep spatiotemporal perturbation modeling according to claim 1, characterized in that: The adaptive spatiotemporal graph convolution module includes an adaptive graph learning unit, which dynamically fuses the physical topology graph structure and the adaptive learning graph structure through learnable weights to capture the implicit spatial dependencies in the road network.

5. The weather-sensing traffic flow prediction method based on deep spatiotemporal perturbation modeling according to claim 4, characterized in that: The adaptive spatiotemporal graph convolution module employs Chebyshev polynomial optimized graph convolution operations, with the order K set to 3.

6. The weather-sensing traffic flow prediction method based on deep spatiotemporal perturbation modeling according to claim 1, characterized in that: The decoupling process in step three is expressed by the formula: Xpure = Xhistory - STDhistory, where Xhistory is the original data and STDhistory is the generated historical spatiotemporal perturbation.

Images